Near-optimal responsive traffic engineering in software defined networks based on deep learning

Abstract

The routing problem for traffic engineering can be solved using different techniques. For example, the problem can be formulated as a linear program (LP) or a mixed-integer linear program (MILP) that requires solving a complex optimization problem. Thus, this approach typically cannot be used for solving a large problem in real time. Alternatively, heuristic algorithms may be devised that, though fast, do not guarantee an optimal decision. This work proposes a novel design of a system that employs a deep learning model trained on optimal decisions to solve the routing problem. The model learns to adapt to traffic dynamics by updating the traffic split ratios to distribute traffic to a few paths between a source and a destination instead of frequently computing a single path for a source and destination pair. This solves the problem of network disturbance and traffic disruption. Specifically, we train two deep learning models: DNN (MLP), which is fully connected layers of neurons, and DNN (LSTM) that consists of a few layers of LSTM neural network and a dense layer. The two models are evaluated in a TE simulator. The system offers two important features of a good traffic engineering system: producing close to optimal traffic engineering results and responding to traffic dynamics in real time. We perform simulations on two topologies, the ATT North America topology, and a 4x4 grid topology. The results show that our proposed system can learn from optimal decisions to attain a responsive traffic engineering system.